NeuronEdge Enclave is now part of Mindpool — the execution boundary in a reference architecture for AI your organization owns.

/security/posture

Security posture at a glance

Two execution tiers. Signed audit. Open source. The security model is the product.

Standard tier

Per-workspace kernel isolation via Firecracker + jailer (chroot, cgroups, seccomp, namespaces). The host operator is trusted.

Confidential profile (preview)

The workspace runs inside an AMD SEV-SNP CVM — operator-excluded, hardware-attested. Key release is gated on a two-layer binding validated to the genuine AMD Milan ARK. The lane stays preview until the exact signed v0.2.0 candidate passes its Azure artifact gate.

Audit

Every workspace action is a signed event in an independently-verifiable Merkle chain. Any party holding an exported manifest can detect edits, broken links, and tail truncation.

Open source

Apache-2.0. The runtime code is public, auditable, and forkable. No security-through-obscurity.

/security/claim-ceiling

What's proven — and what's honestly not

We publish the full ceiling because the honest claim is the product. No overclaim.

Per-workspace kernel isolation (Firecracker + jailer)

Standard profile · lifecycle e2es boot real Firecracker via jailer

Signed, independently-verifiable audit chain

nee audit export + nee audit verify

Azure vTPM SEV-SNP evidence primitive

DCasv5 2026-06-29; 2-layer TPM-Quote binding, validated to the baked Milan ARK

Confidential agent execution (confidential-azure profile)

DCasv5 2026-06-30 round-trip passed; promotion gated on the signed v0.2.0 artifact run

Operator-excluded memory encryption

SEV-SNP: the cloud operator sees ciphertext — within the preview lane above

Attestation-gated key release (sealed snapshots)

CP WASM gate → DEK; seal→unseal byte-identical. Software KEK, not KMS-bound

Confidential snapshot / restore / fork

Not implemented — the profile contract rejects these operations

Runtime supply-chain enforcement (OSV/OPA/CVSS)

Not implemented in standard workspaces; lives in the OpenShell fork

Per-workspace hardware isolation (nested SNP)

Impossible on managed cloud (AMD strips virt extensions)

Guest-code measurement

TCB = host-CVM/OpenHCL launch, not agent code. Tracked follow-on

KMS-hardware-bound key release

Software KEK (no HSM); the AWS KMS backend is mock-tested only, never run against a live account

mTLS runtime↔control-plane

API-key over TLS today

Intel TDX

Needs DCesv5 silicon

Honest ceiling

7 claims are explicitly not built, and 3 more are preview — working, with evidence, but not promoted until the signed release gate passes. We list them because design partners making regulated-enterprise decisions deserve to see the gap. The authoritative record is the runtime's capability ledger and public threat model.

/security/threat-model

The threat model in plain language

what the confidential tier protects against

A compromised cloud operator reading live workspace memory

SEV-SNP memory encryption excludes the cloud operator from live CVM memory — under subpoena they can produce only ciphertext. (Note: this covers live memory only; sealed-snapshot key material is wrapped under a software KEK — see below.)

A compromised host kernel

SEV-SNP integrity protection detects tampering with the CVM's memory pages.

A replayed attestation

The two-layer binding (boot-fixed report + fresh TPM-Quote nonce) means stale or replayed evidence is rejected. Only the live, hardware-anchored key can sign a fresh quote.

what it does not protect against

A compromised agent producing wrong outputs

NeuronEdge Enclave solves execution-boundary safety, not agent alignment. A jailbroken agent contained by Enclave cannot escape the microVM/CVM boundary by design — but it can still produce incorrect or harmful results within its workspace.

Side-channel attacks on the CVM

SEV-SNP has known side-channel limitations (cache timing, etc.). This is a property of the hardware, not of Enclave. We name it honestly.

The paravisor (Azure only)

Azure's OpenHCL paravisor is inside the measured, attested set. On Azure the TCB honestly includes it; the VCEK→ARK signature chain is identical. The bare-metal/GCP (/dev/sev-guest) attestation arm is implemented but not yet silicon-validated — only the Azure OpenHCL arm is verified today.

A runtime insider holding the key-encryption key

Sealed-snapshot data-encryption keys are wrapped under a software KEK held by the runtime/Worker, not hardware-bound to a KMS. The attestation gate is over evidence (policy match), not over key material — an insider with the KEK can unseal. Hardware-binding via AWS KMS is a tracked follow-on.

/security/attestation

Exercised on Azure DCasv5 silicon

The confidential profile is not a slide-deck claim — this round-trip ran on real AMD silicon. It is still preview: the ledger promotes a lane only once the exact signed candidate passes its artifact gate and publishes without a rebuild.

hardwareAzure DCasv5 · Standard_DC2as_v5 · eastus
kernel6.8.0-1058-azure-fde
  1. Nesting block confirmed empirically (/dev/kvm absent, svm cpu flag absent — AMD SEV-SNP strips virtualization extensions from the leaf guest)
  2. OpenShell sandbox spawned in-CVM
  3. Command ran over the SSH control channel
  4. Attestation evidence produced (2-layer binding: boot-fixed report + fresh TPM-Quote nonce)
  5. control plane released the DEK only on that evidence
  6. seal→unseal restored byte-identical plaintext

Run your next agent in a hardware-isolated, governed sandbox.